(PDF) Mesocarnivores in Protected Areas: ecological and anthropogenic determinants of habitat use in northern Kwa-Zulu Natal, South Africa

Mesocarnivores in Protected Areas: ecological and anthropogenic

determinants of habitat use in northern Kwa-Zulu Natal,

South Africa.

Dissertation presented for the degree of

Master of Science

By

Michelle Pretorius

Department of Biological Sciences

Institute for Communities and Wildlife in Africa

University of Cape Town

July 2019

Supervisor: Prof. M Justin O’Riain

Co-Supervisor: Dr Gareth Mann

2

PLAGIARISM DECLARATION

I know the meaning of plagiarism and declare that all of the work in the thesis, save for that

which is properly acknowledged, is my own.

M Pretorius

04/07/2019

3

TABLE OF CONTENTS

ABSTRACT .................................................................................................................................. 4

ACKNOWLEDGEMENTS .............................................................................................................. 5

1 | LITERATURE REVIEW ........................................................................................................... 6

1.1 | Carnivore conservation ................................................................................................................. 6

1.2 | Factors affecting the composition of mesocarnivore communities ............................................ 9

1.3 | Camera traps and occupancy modelling as research tools ........................................................ 12

1.3.1 | Camera trapping theory ...................................................................................................... 12

1.3.2 | Basic camera trap survey designs ....................................................................................... 13

1.3.3 | Occupancy modelling .......................................................................................................... 15

1.4 | Rationale and aims of the study ................................................................................................. 19

2 | METHODS .......................................................................................................................... 21

2.1 | Study areas .................................................................................................................................. 21

2.2 | Study species ............................................................................................................................... 24

2.3 | Camera trap survey design ......................................................................................................... 24

2.4 | Occupancy model covariates ...................................................................................................... 26

2.4.1 | Protected Area (PA) ............................................................................................................ 26

2.4.2 | NDVI ..................................................................................................................................... 26

2.4.3 | Terrain complexity .............................................................................................................. 27

2.4.4 | Distance to PA edge ............................................................................................................ 27

2.4.5 | Domestic dogs ..................................................................................................................... 28

2.4.6 | Apex predators .................................................................................................................... 28

2.5 | Detection covariates .................................................................................................................. 29

2.6 | Multi-species Royle-Nichols occupancy model (RN) ................................................................. 29

2.6.1 | Model framework ............................................................................................................... 30

2.6.2 | Candidate models ............................................................................................................... 32

2.6.3 | Model fit assessment .......................................................................................................... 32

2.6.4 | Derived parameters ............................................................................................................ 33

2.7 | Species richness - Generalized Additive Model ......................................................................... 33

2.8 | Species richness - Variance model .............................................................................................. 34

3 | RESULTS ............................................................................................................................. 36

3.1 | Descriptive results ....................................................................................................................... 36

3.2 | Model selection ........................................................................................................................... 46

3.3 | Mesocarnivore detections and habitat use ................................................................................ 46

3.4 | PA mesocarnivore species richness ............................................................................................ 51

3.5 | Covariates influencing mesocarnivore species richness ............................................................ 52

4 | DISCUSSION ...................................................................................................................... 58

4.1 | Detections ................................................................................................................................... 58

4.2 | Mesocarnivore species richness patterns and species-specific responses ................................ 59

4.3 | PA comparisons ........................................................................................................................... 65

4.4 | Limitations and recommendations ............................................................................................. 66

4.5 | Conclusions ................................................................................................................................. 70

5 | REFERENCES ...................................................................................................................... 72

6 | APPENDIX .......................................................................................................................... 86

4

ABSTRACT

Protected areas (PAs) form the cornerstone for most carnivore conservation strategies.

However, climate change, increased isolation and human pressure along PA boundaries are

together reducing the effectiveness of PAs to conserve carnivores. Mesocarnivores, in

particular, frequently move beyond the boundaries of PAs where they threaten human

livelihoods, and as a result, are often subject to chronic persecution. In South Africa, we know

little about the conservation status of mesocarnivores both within and outside of PAs, as most

research focuses on large, charismatic apex predators. The goal of my study was to leverage

data collected from large carnivore studies to understand variation in mesocarnivore species

richness within PAs. Camera trap surveys were conducted as part of Panthera’s 2015 national

leopard monitoring programme in seven PAs across northern KwaZulu-Natal (KZN), South

Africa. Using a multi-species extension of the Royle-Nichols occupancy model, my study

explored environmental, interspecific and anthropogenic drivers of mesocarnivore habitat

use and species richness. I found a surprisingly low number of detections (N = 356) for all five

mesocarnivore species and considerable variation across PAs. Small PAs with a recent history

of human disturbance supported more mesocarnivore species and at higher relative

abundance. Mesocarnivore species richness was found to decline with increased vegetation

and leopard abundance but increased towards the edge of PAs. Variation in species richness

estimates decreased significantly with vegetation productivity and domestic dog abundance.

Together these results suggest that (1) the edges may provide a refuge for mesocarnivores

from more dominant species, (2) mesocarnivores exhibited resilience/adaptability to human

disturbance, and (3) primary productivity and domestic dog abundance could mediate

mesocarnivore distributions within PAs. My study showed that camera trap data derived from

a single-species survey can be used to make inferences about non-target species to great

success. Current PAs in KZN may not adequately conserve mesocarnivores, and as a result,

emphasis should be placed on coexistence with mesocarnivores in marginal habitat outside

of PAs.

5

ACKNOWLEDGEMENTS

During the last two years I have had the privilege to meet and work with individuals who have

helped me during some stage of my study. It is my pleasure to thank the following individuals

and institutions for their contribution to my thesis.

• First and foremost, I would like to thank my supervisor, Prof Justin O’Riain. I am

extremely grateful for your support, constructive criticism, and general enthusiasm

during the many phases of my study. Also, thank you for generously funding me

through my studies, and I look forward to our future work together.

• To my co-supervisor, Dr Gareth Mann, thank you for providing constructive criticism

and feedback during various critical stages through my project.

• Thank you to the Institute for Communities and Wildlife in Africa and the University of

Cape Town who funded my research.

• Thank you to Panthera for granting me access to their data and allowing me to spend

time in the field gaining experience. Thank you to WildlifeACT and their team of

volunteers and monitors who ran the surveys and conducted initial data capture.

• I would like to thank Matthew Rogan for helping me with numerous data issues, and

his insights and advice regarding my many statistical analyses.

• To all my friends and family, thank you all so much for your general support during my

write-up. In particular, I would like to thank Zoë Woodgate for many brain-storming

chats and chapter swapping. I would also like to thank Tamlyn Engelbrecht, Vincent

Naude, Rebecca Muller and Ruan van Mazijk for always allowing me to vent and

providing a solace away from work.

• Lastly, I would like to thank my parents, Patricia and Tony Pretorius, I truly appreciate

the unconditional love and support you have given me, not only over the last two

years, but my entire academic career. I cannot thank you enough.

6

1 | LITERATURE REVIEW

1.1 | Carnivore conservation

Protected areas (PAs) cover 14.7% of the world’s terrestrial surface (IUCN and UNEP-WCMC,

2019) and are the cornerstone of most carnivore conservation strategies (Hansen and

DeFries, 2007; Caro et al., 2014). They also form a central component of sub-Saharan Africa’s

tourism industry, valued at US$25 billion and provide 2.4% of employment in the region

(WTTC, 2017). Whilst PAs have been acknowledged as important for conserving biodiversity

(Brooks et al., 2006), they are facing a wide variety of threats across a range of spatial scales

which together are reducing their conservation potential, especially for carnivores (Balme,

Slotow and Hunter, 2010; Radeloff et al., 2010; Watson et al., 2014; Santini et al., 2016). These

threats include climate change (Tanner-McAllister, Rhodes and Hockings, 2017), increased

isolation (DeFries et al., 2005), invasive species (De Poorter, 2007) and an increase in

anthropogenic impacts close to PA boundaries (Zommers and Macdonald, 2012), such as

human-livestock-wildlife conflict (Bruner et al., 2001), and the exploitation of natural

resources (Goodman, 2006; Smith et al., 2009; Becker et al., 2013; Caro et al., 2014). These

impacts and their adverse effects on biodiversity are predicted to be disproportionately

greater in developing regions, such as sub-Saharan Africa, largely due to predicted increases

in human populations and the associated developmental changes (Pettorelli et al., 2010).

Carnivores, in particular, have experienced substantial population declines in recent years (Di

Minin et al., 2016). In 2014, meta-analyses on global carnivore conservation revealed that of

the 31 extant terrestrial carnivore species 77% are undergoing continuing population

declines, with 24 of these species under direct threat from human persecution (Ripple et al.,

2014). For example, the Ethiopian wolf (Canis simensis) has experienced dramatic population

declines since 2008 chiefly due to habitat degradation through subsistence farming and

consecutive epizootics of rabies and canine distemper (Marino and Sillero-Zubiri, 2011;

Gordon et al., 2015). Characterised by wide-ranging behaviour, carnivores frequently move

beyond the boundaries of PAs where they pose a threat to human lives, as well as livelihoods,

and are often subject to persecution (Treves and Karanth, 2003). Even within the confines of

PAs, direct poaching can lead to “edge effects”. This may result in total population declines

of carnivores and their prey if balance is not achieved through recruitment (Balme, Slotow

7

and Hunter, 2010; Johnson et al., 2016; Rosenblatt et al., 2016; Carter et al., 2017; Rogan et

al., 2018; van Eeden et al., 2018).

Increased anthropogenic pressure could favour species with greater adaptive plasticity

(Anderson, Panetta and Mitchell-Olds, 2012; Wang et al., 2017). Dietary breadth

1

and

behavioural adaptability allow species to better mediate against environmental changes and

threats (Wong and Candolin, 2015). Large carnivores are often prey specialists and invariably

fulfil the role of apex or keystone species within ecosystems (Ripple et al., 2014), making them

more susceptible to anthropogenic disturbance. By contrast, mesocarnivores

2

are typically

generalist predators, and as such exhibit weaker ecological interactions within the

ecosystems that they live (Roemer, Gompper and Van Valkenburgh, 2009). This makes

mesocarnivores less vulnerable to extinction compared to larger carnivores (Purvis et al.,

2000). However, there has been evidence to suggest that mesocarnivores can provide vital

ecosystem services through seed dispersal (Kurek and Holeksa, 2015; Twigg, Lowe and

Martin, 2016), small mammal control (Ramnanan et al., 2016) and the removal of dead

animals, especially in systems lacking obligate scavengers, such as vultures (Mateo-Tomás et

al., 2015). Thus, reductions in mesocarnivores could be detrimental to overall ecosystem

functioning and human health (Ćirović, Penezić and Krofel, 2016).

The majority of studies (86%) exploring the relationship between apex predators and

mesocarnivores show a strong inverse relationship (Ritchie and Johnson, 2009). Apex

predators suppress mesocarnivore abundance through direct competition and predation

(Wang, Allen and Wilmers, 2015). Declines in apex predator abundance have often been

shown to result in population expansions of mesocarnivore species, known as mesocarnivore

release

3

. This release mechanism has been linked to a negative association between body size

and species richness (Gittleman and Purvis, 1998). This is based on the theory of allometric

ecology, whereby average adult body size (i.e., weight) is strongly related to both physical and

1

Dietary breadth is the range of food items a species can conusume that will maximise the cost/benefit function

of energy per unit of foraging time (Hames and Vickers, 1982). Having a greater diversity in foarging choice can

help buffer against changes in prey perturbations.

2

Carnivore species, also referred to as mesopredators, weighing between 1-15kg (Buskirk, 1999) that hold an

intermediate trophic level (Prugh et al., 2009).

3

An ecological hypothesis whereby decreases in apex predators lead to substantial growth in mesocarnivore

populations (Prugh et al., 2009).

8

behavioural traits (Damuth, 1981). An animal’s body size ultimately determines its relative

energy requirements, influencing both its prey selection and hunting strategies (Carbone and

Gittleman, 2002; Radloff and Du Toit, 2004; Owen-Smith and Mills, 2008).

An animal’s weight has also been linked to a variety of critical ecological variables such as

population density (Johnson, 1999), home range size (Ofstad et al., 2016), dispersal

capabilities (Forero-Medina et al., 2009; Bailey et al., 2018), foraging efficiency (Rizzuto,

Carbone and Pawar, 2018) and fecundity (Allainé et al., 1987). Mesocarnivores, being smaller

than apex predators, typically outnumber them by as much as 9:1 (Carbone and Gittleman,

2002; Roemer, Gompper and Van Valkenburgh, 2009). This ratio becomes more skewed as

mesocarnivores expand to fill vacant niches left by the local disappearance of larger

predators.

The ability of mesocarnivores to rapidly increase in number and exploit both small livestock

on farms and waste within peri-urban areas (Berger, 2006; Prugh et al., 2009; DeVault et al.,

2011; Ripple et al., 2014) has meant they are often assigned the status of pest. For example,

within South Africa, black-backed jackal (Canis mesomelas) and caracal (Caracal caracal) have

dramatically increased in range and abundance following the extirpation of large carnivores

throughout commercial farming areas (Thorn et al., 2012; Drouilly and O’Riain, 2019). While

it is not known what impact apex predators would have had on livestock it is now well

established that mesopredators on farmland are causing significant depredation of sheep

(Ovis aries) and goat (Capra aegagrus hircus), ultimately threatening farmer livelihoods

(Drouilly, Nattrass and O’Riain, 2018; Turpie and Babatopie, 2018). Black-backed jackals in

particular are a concern within rural landscapes given the increased risk of disease

transmission, such as rabies (Butler, Du Toit and Bingham, 2004), and their preference for

livestock over wild prey. In a survey conducted by Badenhorst (2014), six of seven South

African provinces ascribed the majority of their cattle (Bos taurus) depredation to black-

backed jackal. Consequently, black-backed jackals are heavily persecuted on farmland both

by individual hunters and organised culling events (Zimmermann et al., 2009; Turpie and

Babatopie, 2018).

9

1.2 | Factors affecting the composition of mesocarnivore communities

Changeable, complex environments favour generalists as they can rely on several resources

(Clavel, Julliard and Devictor, 2011). Hence, studying generalists requires the examination of

numerous interacting variables. The composition of mesocarnivore communities within PAs

are shaped by various, interacting factors such as habitat requirements, interspecific

relationships, human pressure and PA attributes (Tambling et al., 2018). For example, swift

fox (Vulpes velox) density was found to be negatively influenced by coyote (Canis latrans)

abundance, but this relationship was moderated by basal prey availability and vegetation

structure (Thompson and Gese, 2007).

In general, carnivore distributions within PAs have been shown to be greatly affected by the

presence of permanent water, with increased mesocarnivore occupancy closer to water

sources (Schuette et al., 2013; Rich et al., 2017). Dense vegetation, commonly located around

drainage lines, can provide concealment during hunting and refugia from interspecific

predation (Boydston et al., 2003). Such areas surrounding water sources also offer increased

hunting and scavenging opportunities. For example, black-backed jackals have been observed

killing ungulate calves and a variety of small antelope species next to water sources (Krofel,

2007). African wild cat (Felis silvestris lybica) and civet (Civettictis civetta) also show strong

inverse correlations between occupancy and distance to water (Durant et al., 2010).

Increased habitat variability and structural complexity, i.e., diverse vegetation and terrain,

will also favour mesocarnivores, as it caters to a more generalist niche (Roemer, Gompper

and Van Valkenburgh, 2009; Wilson et al., 2010). For example, caracals have been shown to

prefer rugged terrain that provides both safe sleeping sites and increased ambush

opportunities when hunting (Drouilly et al., 2018).

Prey abundance is thought to directly influence carnivore density (Karanth et al., 2004).

However, the relationship between prey abundance and carnivore presence is complicated

and can be influenced by a multitude of factors, such as interspecific predation and

competition, as well as prey turn-over rates (Fuller, 1996). Flexibility in mesocarnivore diet

enables them to adapt to changes in prey availability (Carbone and Gittleman, 2002),

potentially diminishing the impact of prey abundance on total mesocarnivore density.

10

Caracals and black-backed jackals have been shown to be only marginally affected by changes

in their prey base (Drouilly et al., 2018).

Understanding how local prey populations drive mesocarnivore presence is vital, however,

reliable prey abundance estimates are difficult to obtain. Relative Abundance Indices (RAIs)

can be calculated from animal signs (e.g., tracks or faecal counts), road-kill accounts or

photographs from remote camera traps, whereby the number of detections is used as a proxy

for species abundance. However, these results can be considerably error prone due to the

assumption that all species have an equal probability of being detected over time and space

(Sollmann et al., 2013; Iknayan et al., 2014). For example, carnivores are often assumed to

use roads as a mode of travel (Hines et al., 2010; Poessel et al., 2014; Burton et al., 2015) and

consequently many studies exploring carnivore occupancy or abundance survey along road

networks to increase detection probability. However, there is evidence that prey species are

indifferent to roads, or actively avoid them, and therefore, using RAIs derived from a road

based survey would lead to inaccurate assessments of prey availability (Harmsen et al., 2010;

Mann, O’Riain and Parker, 2015). It is thus difficult for any one method (e.g., camera traps or

tracks) or survey design (paths, roads or random) to provide accurate and cost-effective

estimates of both predator and prey species abundance.

Interspecific interactions play a major role in determining carnivore population density

(Fuller, 1996). Interspecific predation and competition can lead to changes in species

abundances within PAs, as a result of elevated levels of aggression and kleptoparasitism. On

average, an African carnivore will experience exploitative competition with 22.4 other species

(Caro and Stoner, 2003). Mesocarnivore species are particularly vulnerable to

kleptoparasitism with 13 different species stealing food from black-backed jackals, 10 from

serval (Leptailurus serval), 9 from caracal and 3 from side-striped jackal (Canis adustus; Caro

and Stoner, 2003). Subordinate carnivores, such as the Altai mountain weasel (Mustela

altaica), have been shown to spatially and temporally shift their behaviour to avoid its more

dominant competitors, stone martens (Martes foina) and red foxes (Vulpes vulpes; Bischof et

al., 2014). Due to the inherent functional diversity present in the mesocarnivore guild, they

can exploit a broad prey base (Roemer, Gompper and Van Valkenburgh, 2009; Rich et al.,

2017), and as a result, can exhibit a variety of responses to similar external pressures.

11

Caracals, for example, have been observed to be more successful in the absence of apex

predators than black-backed jackals (Drouilly et al., 2018). Therefore, the relationship

between large and mesocarnivores may reflect a complex balance of risk-avoidance and

energy requirements, all of which may be influenced by anthropogenic disturbances.

Human presence on the boundaries of PAs can be detrimental to the overall health of the PA

(Hansen and DeFries, 2007; Radeloff et al., 2010). Reduction and/or degradation (from direct

habitat loss, or noise and light pollution) in the areas around a PA can greatly reduce its

functional size. Subsequently, as core area size is reduced, animals at higher trophic levels,

such as large apex predators, are the first to become locally extinct (Ripple et al., 2014) . This

can lead to trophic cascades or mesocarnivore release within a PA (Ritchie and Johnson, 2009;

Newsome et al., 2017). Human presence can also lead to the destruction of potential corridors

between PAs, or ephemeral lands, restricting dispersal and gene flow between protected

populations. Finally, increased exposure to humans at the edge of PAs can lead to population

sinks, caused by legal offtake through hunting, illegal poaching, the introduction of invasive

species and diseases, or increased human-wildlife conflict. Radeloff et al. (2010) showed that

housing growth within a 1km buffer of PA boundaries exerted a direct influence on the wildlife

within.

The introduction of domestic animals and invasive species by local communities has a large

negative impact on carnivore species (Hughes and Macdonald, 2013; Zapata-Ríos and Branch,

2016). The proximity of livestock to PA boundaries increases retaliatory killing (Berger, 2006;

du Plessis et al., 2015), as livestock provide a plentiful and easy prey resource for carnivores.

Domestic dogs (Canis familiaris) pose a significant threat to native carnivores by acting as

competitors, predators and disease vectors (Young et al., 2011; Silva-Rodríguez and Sieving,

2012; Gompper, 2013; Zapata-Ríos and Branch, 2018). Domestic dogs are found in higher

densities in more human-dominated areas (Odell and Knight, 2001; Ordeñana et al., 2010)

and where agricultural land borders PAs. The presence of domestic dogs in PAs and their

associated threats to native species are most significant at the borders, showing a decreasing

trend to the interior of the PA (Torres and Prado, 2011). Therefore, dogs could exacerbate

edge effects associated with the peripheries of PAs (Revilla, Palomares and Delibes, 2006).

12

All of these anthropogenic influences may manifest as edge effects; increased mortality rates

close to or beyond the reserve boundaries, causing these peripheral areas to become

population sinks (Woodroffe & Ginsberg 1998). Edge-dwelling species are positively affected

by changes in the core:edge ratio of a PA. Mesocarnivores might not be effected by edge

effects to the same extent as larger carnivores due to their ecological plasticity (Purvis et al.,

2000). However, mesocarnivore responses have been shown to vary depending on species

sensitivity to fragmentation and anthropogenic factors (Baker and Leberg, 2018). As

mesocarnivores constitute a large portion of road kill and are heavily persecuted due to

livestock predation or perceived rabies threats (Noss, 1998), utilising available edge space

may be harmful to mesocarnivore abundance due to an increased chance of hostile contact

with humans (Crooks, 2002).

1.3 | Camera traps and occupancy modelling as research tools

To improve large scale predator management, it is vital that we understand how carnivore

species of all sizes and trophic levels interact with their environment. This requires long-term

monitoring of populations, at varying spatial and temporal scales, across a range of climatic

and land use gradients.

1.3.1 | Camera trapping theory

Camera trapping has emerged as a popular, non-invasive and cost effective method for

monitoring wildlife presence, abundance, activity patterns and density (O’Connor et al.,

2017). It allows for longer monitoring periods and larger-scale research than traditional

surveying techniques (e.g., scat surveys, line transects or GPS collars), creating the

opportunity to study rare, elusive species (Caravaggi et al., 2017).

The advantages of camera trapping are not limited to field work but expand into the data

extraction process. Using camera trapping as a means of data collection allows for post-study

data verification and analyses by independent observers (Newey et al., 2015; Caravaggi et al.,

2017). This enables datasets to be inherited and subsequently re-evaluated under new

methodologies or used to explore alternative hypotheses. Additionally, processing time and

13

cost can be greatly reduced by using citizen scientists (non-professional volunteers) for

species identification (Kosmala et al., 2016). Artificial intelligence (AI) is also being used to

assist in wildlife monitoring (Kwok, 2019). Together these advances have facilitated larger and

longer running camera trap surveys as data processing is no longer restricted by researchers’

available time or funds. AI can facilitate the quick accumulation and analyses of large datasets

ultimately allowing for more robust and timeous conclusions.

The most common use of camera trap data is estimating abundance or density of individually-

identifiable species (e.g., tigers (O’Brien, Kinnaird and Wibisono, 2003; Wang and Macdonald,

2009) or leopards (Balme, Slotow and Hunter, 2010)) through spatial capture-recapture

modelling. Yet the versatility of camera trapping has allowed for a wide range of applications

such as: biodiversity assessments (Ahumada et al., 2011), species discoveries (Rovero et al.,

2008) and behavioural studies. Behaviours that have been monitored with camera traps

include anti-predator responses (Carthey and Banks, 2015), denning (Miller et al., 2017),

foraging (Delgado-V et al., 2011), resource portioning (Edwards, Gange and Wiesel, 2015),

social behaviour (Leuchtenberger et al., 2014) and temporal avoidance (Romero-Muñoz et

al., 2010). Finally, incorporating presence-absence data from camera traps into occupancy

modelling frameworks has facilitated investigations into the drivers of unmarked species

(species with unidentifiable characteristics) distributions and species richness (Kolowski and

Forrester, 2017).

1.3.2 | Basic camera trap survey designs

Camera trap survey design (i.e., camera locations, numbers, spacing and length of

deployment) should be linked to the overall study objectives (Burton et al., 2015). Studies

may prioritise landscape sampling over detection rates, and therefore, design the camera

survey so as to accurately sample the available landscape (Kolowski and Forrester, 2017;

O’Connor et al., 2017). Spacing within a survey is determined by placing a grid over the study

area and apportioning camera stations within each grid (Figure 1.1). Camera traps can be

spaced at random in a landscape using a simple random design (e.g., allocating camera

locations at random coordinates within a grid) or a systematic random design (e.g., locations

are arranged in a regular pattern at equidistance from each other). A clustered camera design

14

can also be used when accessibility within a study area is limited (Figure 1.1). Grid cell size

can be determined based on the number of cameras available or on individual home range

size of the target species. Some camera trap surveys also bait or use scent lures to increase

detection probability (du Preez, Loveridge and Macdonald, 2014).

Figure 1.1 Basic sampling designs for camera trap surveys taken from Wearn and Glover-

Kapfer (2017)

A large percentage of camera trap surveys (54.8%) use capture-recapture methods based on

a targeted probabilistic design (Burton et al., 2015). This is where cameras within each grid

cell are purposefully placed along a corridor of animal movement, such as rivers, roads or

trails, so as to maximise the probability of detecting the target species (Karanth et al., 2004;

Shannon, Lewis and Gerber, 2014). This limits the usefulness of the data for other purposes,

such as in biodiversity estimations or for monitoring non-target species using RAIs (Wearn

and Glover-Kapfer, 2017). Also, using such “optimal” camera locations can potentially cause

a bias in detection rates, especially when multiple species are being studied (O’Brien, 2011;

Swann, Kawanishi and Palmer, 2011; Wearn et al., 2017). However, robust statistical

methods, such as occupancy modelling, can mitigate detection biases created by unsuitable

camera trapping designs (MacKenzie et al., 2006; Gould et al., 2019).

The success of a camera trapping survey is heavily dependent on individual camera reliability.

Camera failure does occur, such as when batteries malfunction, or difficult climatic conditions

lead to increased non-animal trigger events which can rapidly saturate a camera’s memory or

lead to premature battery failure (Swann, Kawanishi and Palmer, 2011). Fortunately, many of

Simple random Systematic random Clustered

15

these technical failures can be accounted for either in the survey design, i.e., having multiple

checks throughout the survey to reduce prolonged camera faults, or during post hoc data

cleaning, such as correcting erroneous date/time stamps (Shannon, Lewis and Gerber, 2014).

Another challenge associated with camera trapping is inconsistency in terminology and

underreporting of survey methodology in research papers, preventing reproducibility (Meek

et al., 2014).

Sampling error is a common problem with any wildlife surveying method, particularly with

regards to imperfect detection. Although a species is present in the sample unit, an individual

may not be detected by the camera (MacKenzie et al., 2006). Imperfect detection can occur

on multiple levels. Firstly, the animal can use the area within the detection zone but not

trigger a capture event, due to its body size, movement speed or wariness for foreign objects

(Caravaggi et al., 2017). Secondly, the animal can use the wider area around the camera but

not enter the detection zone, i.e., camera traps can only detect movement within a certain

radius of the camera’s sensor (Burton et al., 2015). Finally, an animal may also become

temporally unavailable for detection due to its episodic or mobile nature, or other individuals

may immigrate into the study area or be recruited through birth. These challenges can be

addressed through repeated surveys and/or by incorporating camera placement features,

and other detection variables, into further modelling exercises to reduce the probability of

false absences and prevent misleading conclusions on species distribution (MacKenzie et al.,

2006; Royle and Dorazio, 2008; Mordecai et al., 2011).

1.3.3 | Occupancy modelling

Camera trap data is particularly useful for estimating occupancy. Occupancy modelling offers

great flexibility, answering a variety of complex ecological questions with relatively simple

sampling designs, whilst accounting for imperfect detection. It has been used successfully to

explore a variety of different concepts, such as geographic range (Bled, Nichols and Altwegg,

2013), ecological niche partitioning (Schuette et al., 2013), anthropogenic effects on

populations and communities (Hames et al., 2002) and resource use (Manly et al., 2002). The

flexibility and ease of occupancy modelling has also facilitated multi-species studies which can

16

explore species interactions (Steinmetz, Seuaturien and Chutipong, 2013; Rota et al., 2016;

Drouilly, Clark and O’Riain, 2018) and community dynamics (Burton et al., 2012).

The basis of occupancy modelling is to explore what factors determine the proportion of sites

occupied by a species. If one assumes that species occupancy (

!

) is independent of any

variables, the naïve estimate would be the number of sites occupied over the total number

of sites surveyed, i.e.,

! = #$$%&'()

*#*+,

(MacKenzie et al., 2006). However, this fails to account

for both detection error and drivers of species occupancy. A robust occupancy model must

incorporate covariates as well as detection probability (Figure 1.2; MacKenzie et al., 2006;

Guillera-Arroita, 2017). Most occupancy models use a hierarchical model structure

integrating two processes (Figure 1.2A): a system constituting the underlying biological

system (i.e., species occupancy and the associated covariates) and the detection process

(Tobler et al., 2015; Guillera-Arroita, 2017). This general two-level hierarchical model

accounts for the fact that the biological system is only partially available for observation due

to imperfect detection (King, 2014).

Describing the biological system usually takes the form of a logistic regression which relates

the probability that a species is present at a site to the associated environmental predictors

through a logit link function (Figure 1.2B). The parameters related to the probability of

occupancy are conditional on the underlying occupancy state. Assuming no false positive

detection (Type I error), a species cannot be detected at a site it does not occupy, while at a

site that is occupied the species will be detected based on the given detection probability

(Mackenzie et al., 2003; MacKenzie et al., 2006). Model fitting can be implemented within the

frequentist (maximum-likelihood) or Bayesian frameworks.

17

Figure 1.2 Breakdown of the data structure and modelling needs of single-species, single-

season occupancy models, adapted from Guillera-Arroita (2017). A) An occupancy model has

two components: one that describes the relationship between the chosen covariates and

species occupancy, and one that describes how the observed occupancy pattern can be

influenced by detection. B) An example of how to construct an occupancy model. The

detection history is in the form of binary records

-'.

. The presence-absence of a species at a

site

/'

is modelled using a logistic regression, as a function of two site level covariates:

0

and

1

. The probability of detection

2'.

is modelled through a second logistic regression, as a

function of two covariates: site-specific

3

, and survey-specific

4

.

The rise of occupancy modelling has allowed for camera trapping surveys to be used as a

surrogate method for estimating abundance (Mackenzie and Royle, 2005; Ahumada et al.,

2011; Mordecai et al., 2011). The assumptions of an occupancy model include (Mackenzie et

al., 2003):

1. Occupancy state is “closed”. That is, the species of interest is present within the site

for the duration of the survey and its occupancy does not change during the course of

the sampling period.

2. Sites are independent. Detection of a species at one site is independent of detecting

the species at other sites.

3. No unexplained heterogeneity in occupancy. The probability of occupancy is constant

across sites or if differences in probability do occur, they can be explained by the

selected covariates included in the model.

4. No unexplained heterogeneity in detection. This is similar to assumption 3 whereby

any variation in detectability between sites must be accounted for with selected

covariates in the model.

A B

Detection matrix

18

New approaches to occupancy modelling are regularly being developed that allow for

increased flexibility in study design (Altwegg and Nichols, 2019), relaxation of model

assumptions (Gould et al., 2019), the evaluation of complex relationships (e.g., species

interactions; Rota et al., 2016), and the ability to account for additional sources of bias (e.g.,

“false positive” detections associated with species misidentification; Ferguson, Conroy and

Hepinstall-Cymerman, 2015). The closure assumption, assumption 1 above, is often violated

with wide-ranging, territorial species that move in and out of the sample unit. However,

recent studies have shown that occupancy models can still be an effective tool for studying

the distribution of highly mobile species even when the assumption of geographic closure is

ignored (Gould et al., 2019). Although violating this assumption may not bias results, the

produced estimates should be interpreted as the probability of “use” rather than occupancy

(Kendall and White, 2009). Non-independence in sites can arise when sample sites are located

too close to one another, allowing an individual to be detected at multiple sites

simultaneously. This results in overdispersion as the true number of independent sites is

smaller than the number of sites sampled and can result in overestimated abundances,

potentially leading to inappropriate management decisions (MacKenzie and Bailey, 2004;

Martin et al., 2011). Goodness-of-fit assessments have the power to detect and adjust for this

dispersion (MacKenzie et al., 2006).

The last two assumptions imply that variation in occupancy and detection probability is

appropriately modelled with the chosen covariates, i.e., there is no unmodeled variation. This

is a typical situation in occupancy modelling, especially when analysing historical or inherited

data, where the required covariates were not measured at the time. For example, carnivore

occupancy is likely a function of local prey abundance or density, however, without a priori

knowledge on these numbers this covariate cannot be confidently included in the model

(Gerber et al., 2009). Coarse-scale proxy variables can be used, but this may not be available

or poorly represent the true distribution of the desired covariates. This unobserved detection

heterogeneity can be addressed using finite mixtures, where multiple finite detection

probabilities are considered, or by incorporating a random effect, where detection

probabilities are treated as a probability distribution with a mean

5

and standard deviation

6

(Mackenzie et al., 2003; Royle and Nichols, 2003; MacKenzie et al., 2006; Gerber et al., 2009).

19

1.4 | Rationale and aims of the study

Mesocarnivores are poorly studied (du Plessis et al., 2015) with most research focused on

large, charismatic apex predators. They occur across ecosystems with a range of trophic

pressures; from those in which large predators still persist, to others that have been

completely transformed through urbanisation or agriculture (Tambling et al., 2018). In South

Africa, the trophic status of mesocarnivores is largely unknown and case-specific.

Consequently, we know little about the conservation status of mesocarnivores both within

and outside of PAs. Anecdotal evidence, such as the disappearance of black-backed jackals

from Hluhluwe-Imfolozi National Park in KwaZulu-Natal (KZN) and their subsequent

reintroductions and repeated demise in the 1990s (Somers et al., 2017), suggest that there

are serious unknown threats to mesocarnivores.

Within PAs, research opportunities on mesocarnivores are limited largely because they are

not seen to be as ecologically important as their larger counterparts (Roemer, Gompper and

Van Valkenburgh, 2009) and are less likely to attract funding or interest from tourists. To

circumvent this, one can leverage research efforts on larger carnivores to study

mesopredators within PAs. Thus, a camera trap survey designed to monitor apex predator

abundance can provide useful data on mesocarnivores. Although these surveys may not be

optimized for mesocarnivore research, largely due to camera placement, the data are still

potentially valuable for exploring mesocarnivore presence within PAs.

In this study I use data collected as part of a large scale project on leopard density within PAs

in KZN to explore variation in mesocarnivore presence and richness. I selected seven PAs that

range in size, management history and surrounding land-use. I aim to test the hypothesis that

mesocarnivore habitat use and richness within PAs is influenced by a broad range of

environmental and anthropogenic variables. These variables include the presence of larger

predator species, vegetation characteristics, terrain complexity and human disturbance

surrounding the PA. Collecting data over a variety of PAs allowed me to investigate possible

reasons for observed mesocarnivore population distributions within PAs. Additionally, gaining

a baseline of information on the current population status can allow for better management

20

of mesocarnivore species in these areas, especially under continued habitat change and

persecution.

21

2 | METHODS

I adopt occupancy nomenclature, in which ”site” refers to the selected PA in which the surveys

occurred. “Survey” defines a continuous primary sampling period within the site in a given

year. The survey period is subdivided into a number of secondary sampling ”occasions” over

which sampling is replicated, and finally, ”station” defines the location of a pair of camera

traps.

2.1 | Study areas

Camera surveys were conducted as part of Panthera’s national leopard monitoring

programme, established to monitor leopard population status across South Africa. The

programme started in 2013 and has expanded through time such that a total of 20 PAs were

surveyed in 2018. This study utilized a subset of the camera trap data collected from seven

PAs distributed across northern KwaZulu-Natal (KZN), South Africa in 2015 (Figure 2.1). All

seven sites are classified as “Nature Reserves” (Protected Areas Act 57, 2003). These sites

included the Eastern Shores Section of iSimangaliso Wetland Park (Eastern Shores), Hluhluwe-

Imfolozi Park (HiP), Ithala Game Reserve, Somkhanda Game Reserve, Tembe Elephant Park,

uMkhuze Game Reserve and Zululand Rhino Reserve (ZRR). Management authorities varied

between sites (Table 2.1) and included provincial (n = 5), community (n = 1) and private (n =

1) PAs. Sites were on average 377

±

236 km2 in size and collectively covered 2643 km2 (Table

2.1).

Zululand Lowveld vegetation was the dominant vegetation type within three of the study

sites, namely HiP, Somkhanda and ZRR. Dominant vegetation in the other sites included

Tembe Sandy Bushveld (Tembe), Western Maputaland Clay Bushveld (uMkhuze), Ithala

Quarzite Sourveld (Ithala) and Subtropical Freshwater Wetlands (Eastern Shores) (Table A3).

22

Figure 2.1. Land-use map of the study area in northern KZN, South Africa. Camera stations (+) were located in seven study sites/PAs: (1) Eastern Shores (n = 41

camera stations) sampled between September and November 2015, (2) HiP (n = 46) sampled between May and June 2015, (3) Ithala (n = 31) between March

and June 2015, (4) Somkhanda (n = 41) between January and March 2015, (5) Tembe (n = 32) between July and August 2015, (6) uMkhuze (n = 40) between

June and July 2015, and (7) ZRR (n = 40) between July and September 2015. Numbers correspond with Table 2.1.

Swaziland

South Africa

Mozambique

+

Urban

Agriculture

Degraded land

Water

Other

Study site

Camera stations

Rivers

National boundary

Legend

N

6

1

2

34

7

5

Durban

0 25 50km

23

Table 2.1 The names and size of the seven study sites located in northern KZN, South Africa. Camera surveys were conducted as part of Panthera’s

2015 leopard monitoring programme. “Site” refers to the selected PA in which the survey occurred, ”station” defines the location of a pair of

camera traps and “occasions” are the number of secondary sampling periods. Bold numbers correspond with Figure 2.1.

Site name

Area (km2)

No. of

stations

Start date

End date

No. of

occasions*

Season

Management

authority

1

Eastern Shores Section

of iSimangaliso Wetland

Park

264

41

25/09/2015

08/11/2015

1832

Dry/Wet

Provincial

2

Hluhluwe-Imfolozi Park

904

46

01/05/2015

14/06/2015

2045

Dry

Provincial

3

Ithala Game Reserve

292

31

29/03/2015

12/05/2015

1349

Wet

Provincial

4

Somkhanda Game

Reserve

313

41

30/01/2015

15/03/2015

1737

Wet

Community

5

Tembe Elephant Park

299

32

12/07/2015

25/08/2015

1392

Dry

Provincial

6

uMkhuze Game Reserve

355

40

02/06/2015

16/07/2015

1759

Dry

Provincial

7

Zululand Rhino Reserve

216

40

29/07/2015

11/09/2015

1793

Dry

Private

*camera occasions were calculated by summing all days the camera stations were active

24

2.2 | Study species

Table 2.2 The common and scientific names of the five mesocarnivore species included in my

study in addition to the average adult body mass, calculated from Hunter and Barrett (2011),

IUCN conservation status taken from IUCN (2019) and the South African status taken from

Friedmann and Daly (2004). LC - Least concern and NT - Near threatened.

Common name

Scientific name

Average adult body

mass (kg)

IUCN

status

South African

status

Caracal

Caracal caracal

14

LC

Side-striped jackal

Canis adustus

13

LC

NT

Black-backed jackal

Canis mesomelas

13

LC

Honey badger

Mellivora capensis

11

LC

NT

Serval

Leptailurus serval

11

LC

NT

All five mesocarnivore species included in the study (Table 2.2) are thought to be widespread

throughout South Africa. Black-backed jackals and caracals are ubiquitous in agricultural areas

and, as a result are commonly persecuted. The other three species are also associated with

livestock predation, though to a lesser degree (Kerley et al., 2018). Therefore, understanding

the presence of these agricultural ‘pest’ species in PAs is important for their overall

management (Turpie and Babatopie, 2018).

2.3 | Camera trap survey design

Within each site, 31-46 paired camera trap stations were deployed for approximately 45 ± 2

days (Table 2.1, Figure 2.1). Stations consisted of a pair of unbaited, motion-triggered

PantheraCam V-series camera traps (Figure 2.2; Olliff et al., 2014) positioned approximately

40 cm above the ground on trees or steel poles. Stations were spaced 1-3 km apart so as to

maximise the probability of detecting Panthera’s target species, leopards (Tobler and Powell,

2013). This distancing has also been shown to be appropriate for monitoring multiple

medium- to large-sized mammals in forest and savanna/grassland systems (O’Brien et al.,

2010). Cameras were set alongside publicly-accessible roads, management tracks, drainage

lines and well-used animal paths to optimise detection probabilities (Figure 2.2 and Figure

2.3). Cameras were placed opposite each other on either side of the track but not facing each

other directly (offset by ~2 m) to avoid camera flash reflection. Each station was treated as a

single data point by combining detections from the two opposing cameras using the

time/date stamps recorded (e.g., if a caracal was recorded by both cameras and shared a

25

date/time stamp, it was recorded as a single detection). Independent captures were defined

as photo events separated by

!

30 minutes unless different individuals could be distinguished.

Finally, to minimise false detections, vegetation around each individual camera that might

obstruct the camera’s field of view was cleared. Camera traps were not moved during the

individual surveys. Cameras were checked every 7-10 days to replace batteries, download

images and perform any other maintenance tasks that were required. Camera settings

included medium trigger sensitivity, with a flash distance of ~5 m (Figure 2.3). When flash was

employed for nocturnal captures, a single image was taken per trigger with an 8 second delay

between each image. When flash was not used, 3 daylight images were captured per trigger

with a 0.3 second delay between each image.

A detection history was then created for each individual species based on the date and

location of individual captures. For each site, the observed data consisted of an

" # $

matrix

where

"

was the number of stations and

%$

was the number of occasions. A camera station

was considered to be active if at least one of the two cameras was operational. For each

occasion, the species was either registered as detected (“1”) or not detected (“0”). The

occasion matrix was later pooled into 5-day occasion periods to reduce zero-inflation and

improve model fit.

Figure 2.2 PantheraCam V-series camera trap (insert) and mounted on metal pole along an

animal track.

26

Figure 2.3 Schematic showing the detection zone of a PantheraCam V-series camera trap

mounted on a tree and placed along an animal trail.

2.4 | Occupancy model covariates

2.4.1 | Protected Area (PA)

There are many aspects of a PA that are difficult to measure, or are unknown, such as its land-

use history, past and current management, or disease outbreaks. Therefore, PA was included

as a covariate to try and incorporate some of this variation into the final model.

2.4.2 | NDVI

Increased habitat closure was thought to increase mesocarnivore habitat use, as it provides

refugia for smaller, less-dominant species (Roemer, Gompper and Van Valkenburgh, 2009;

Wilson et al., 2010). Vegetation primary productivity was estimated using the Normalised

Difference Vegetation Index (NDVI), computed as:

&'( )*+,

&'( -*+,

where

&'(

is the amount of near-infrared light and

*+,

is the amount of red light reflected

by a surface and measured by a satellite sensor (Pettorelli et al., 2011). Plant material

generally has high visible light absorption and high near-infrared reflectance, thus resulting

in positive NDVI values that represent photosynthetic activity and canopy structure (Pettorelli

et al., 2005, 2010). NDVI has been found to decrease as vegetation becomes more open (e.g.,

forest to grassland; Martinuzzi et al., 2008). NDVI observations were acquired from the Tera

MODIS MOD13A1 Version 6 dataset (Huete et al., 2002). NDVI was sampled on a pixel scale

every 16 days at 500 m resolution. NDVI within a 1 km buffer area around each station was

Tra i l

Tre e

Camera

Detection

zone

5m

1m

27

averaged and values were limited to periods encapsulating the respective survey period of

approximately 45 days (Table 2.1).

2.4.3 | Terrain complexity

Water availability is difficult to measure as many satellite or aerial methods cannot account

for all forms of “available water”, such as many temporary water sources, soil moisture and

groundwater (Rockström et al., 2009; Gerten et al., 2011). Multiple surveys in this study were

conducted over wet seasons where temporary water sources would have been prevalent.

Terrain complexity can reflect hydrological profiles and also influence availability of refuge

and foraging diversity, ultimately impacting prey densities (Berryman et al., 2015). Therefore,

it was thought that terrain complexity may be a more important variable than distance to

water. Terrain complexity has also been shown to be important in estimating the occupancy

of black-backed jackals and caracals (Drouilly, Clark and O’Riain, 2018); therefore, I

hypothesised that mesocarnivore habitat use would increase in areas with greater terrain

complexity. Terrain complexity is a convoluted factor, and difficult to quantify; thus, the

Terrain Ruggedness Index (TRI) was used as a proxy variable (Pitman et al., 2017). TRI was

derived from 30 m resolution Shuttle Radar Topography Mission (SRTM) elevation data

(USGS, 2014). Each pixel was rescaled to 500 m and calculated as the square root of the

summed squared difference of a pixel and its eight neighbours. TRI at each station was then

calculated as the average TRI within a 1 km buffer area around each station.

2.4.4 | Distance to PA edge

Reduction and/or degradation due to human presence around PAs can greatly reduce its

overall health and functional size. Additionally, these edge effects can lead to population sinks

from illegal poaching, introduction of invasive species and diseases, or increased human-

wildlife conflict (Woodroffe and Ginsberg, 1998; Massey, King and Foufopoulos, 2014).

Therefore, I hypothesised that mesocarnivore habitat use and species richness would

decrease closer to the edges of PAs. Distance to the edge of the PA was used as a proxy for

these hypothesised edge effects. Spatial polygons for each PA were aggregated from land

owners, PA management and the World Database on Protected Areas (IUCN and UNEP-

28

WCMC, 2019). Distance to the edge was then calculated as the linear distance from a camera

station to the nearest point outside of the PA.

2.4.5 | Domestic dogs

As domestic dogs generally occur in close proximity to human-dominated areas (Silva-

Rodríguez and Sieving, 2012), it was thought that the presence of domestic dogs within a PA

would be a good proxy for human disturbance bordering the PA , as well as the permeability

of the surrounding fence structure. Additionally, dogs are often used in bushmeat hunting

(Jachmann, 2008; Grey-ross, Downs and Kirkman, 2010; Lindsey et al., 2011), and thus could

also reflect poaching pressure within the PAs. It was hypothesised that both mesocarnivore

habitat use and richness would decrease with increased domestic dog abundance due to

direct competition and disease risk (Zapata-Ríos and Branch, 2018). Domestic dog relative

abundance (hereafter referred to as “dogs”) was estimated using Relative Abundance Indices

(RAI), where total independent (

!

30 min) dog detections were summed for each station,

standardised by effort, i.e., the number of days the station was active/operational, and

multiplied by 100.

2.4.6 | Apex predators

It was thought that interspecific predation and competition would play a major role in

determining mesocarnivore habitat use and richness (Fuller, 1996). Large carnivores, such as

leopard, lion (Panthera leo) and spotted hyaena (Crocuta Crocuta), often prey upon smaller

carnivores such as caracal, honey badger (Mellivora capensis) and jackal species, and can

significantly shape mesocarnivore communities through direct mortality and induced

avoidance behaviour (Hayward et al., 2006; Ramesh, Kalle and Downs, 2017b). I hypothesised

that mesocarnivores would be more successful, that is have increased habitat use and species

richness, with reduced apex predator abundance (Drouilly et al., 2018). Apex predator RAI

was calculated in the same way as domestic dogs where the total number of independent

apex carnivore detections were summed across species (leopard, lion and spotted hyaena)

and stations, standardised by effort, and multiplied by 100.

29

2.5 | Detection covariates

Positioning of cameras by landscape features such as roads, animal tracks and/or river beds,

hereafter referred to as “trail”, was included as a detection factor as this may have favoured

the detection of some mesocarnivore species over others. Due to the variety of dominant

vegetation located within each study site (Table A3), PA was also included as a detection

covariate.

2.6 | Multi-species Royle-Nichols occupancy model (RN)

As previously mentioned, occupancy models provide an estimate of occupancy (

.

), i.e., the

probability that the study site is occupied by a particular species during the survey period

(Mackenzie et al., 2002). As my study dealt with highly-mobile species, with male black-

backed jackals having an average home range of 18.2 km2 within a KZN PA (Rowe-Rowe, 1982)

and 288.5 km2 for male caracals in KZN agricultural landscape (Ramesh, Kalle and Downs,

2017a), sampling unit closure could not be assumed (assumptions outline on page 17), and

thus the model outputs were interpreted as the probability that a species “uses” a site, rather

than the traditional occupancy probability. Despite violating the assumption that sampling

units are independent, occupancy models are still an effective tool in estimating habitat use

of highly-mobile species (Gould et al., 2019). For the model to accurately estimate habitat

use, imperfect detection and covariates thought to influence both use and detection must be

incorporated (Reilly et al., 2017).

Tobler et al. (2015) developed a multi-species extension of the Royle-Nichols (RN) occupancy

model (Royle and Nichols, 2003) that utilizes camera trap detection data to monitor a variety

of species compositions and occupancy over time. The basic single-species Royle-Nichols

model estimates occupancy rate or the proportion of area used as a function of species

abundance. The model assumes that variations in abundance will induce variation in

detection probability (Royle and Nichols, 2003). Therefore, their method estimates species

abundance from repeat observations of the presence-absence of a species, without requiring

individual identification. The multi-species extension of this model allows for multiple species

abundances to be estimated and allows multiple study areas to be compared, while still

accounting for species-specific differences in detection. I adapted this multi-species RN model

30

to analyse how different PAs, as well as human disturbance and environmental covariates

affect species-level and guild-level mesocarnivore habitat use.

2.6.1 | Model framework

The theoretical background of my model follows Tobler et al. (2015) closely, while model

specifics were obtained from Li, Bleisch and Jiang (2018).

/01

was defined as a latent binary

variable that indicated whether the species of interest

2

was present

3/01 4 56

or absent

3/01 4 76

in the PA

8

. It was assumed that

/01

was a Bernoulli random variable and governed

by the hyperparameter

4

90

, where

90

is a rate between 1 and 0 indicating the proportion of

sites where species

2

was present.

/01%:%;+*<=>8823906

90

can also be thought of as habitat use at the PA level, this is important as it allows some

species to be completely absent from certain PAs. The observed occurrence state of species

2

at camera station

?

was defined by the binary variable

@0A

where:

@0A%:%;+*<=>882

B

/01 #.0A

C,

where

.0A

is the probability that species

2

uses the area around station

?

.

The RN model estimates the abundance distribution of species

2

using the probability of

habitat use, such that:

.0A 4DE

B

F0A G 7

C

4 5 )HIJ%3)K0A6

Where the abundance of species

2

at station

?

3F0A6

is modelled as a random Poisson variable

K0A

;

F0A:L=2MM=<3K0A6

,

It was hypothesized that mesocarnivore habitat use would vary with ecological and

anthropogenic covariates (COV: NDVI, TRI, distance to PA edge, domestic dog and apex

predator abundance) and across all PAs.

4

A hyperparameter is a parameter whose value is based off a prior distribution (Riggelsen, 2006). For example,

if the variance parameter

NO

has a uniform prior of (0,

%P

), then

NO

is a parameter (in the distribution of the data)

and

P

is a hyperparameter, as it is not based on observed data.

31

Thus, the expected abundances

3K0A6

were calculated as:

QRS

B

K0A

C

4 TU0 -

V

TW0XYZW-T[0L\A

]^_

W

where

`

is the index of the five covariates. Individual detection probabilities varied with PA

and trail:

8=a2b

B

c0A

C

4%dU0 -%de0b*F28A-%dO0L\A

All continuous covariates were standardized to have a mean of zero and a standard deviation

of one.

Following Reilly et al. (2017) and Li, Bleisch and Jiang (2018), species-level parameters were

modelled as random variables drawn from a normal distribution with a mean of zero and

variance

3fO6

described by community hyperparameters, e.g.,

Te0:&=*gF837hfO6

. This was

done to improve parameter estimations for scarcely detected species. Following Rich et al.

(2016) all species were pooled into a “community”, with all species richness estimates being

derived from the community model with community-level hyperparameters.

Modelling was carried out in a Bayesian framework using BUGS (Bayesian inference Using

Gibbs Sampling) language and run in JAGS (Just Another Gibbs Sampler) software (Plummer,

2003; version 3.4.0) with the R package R2Jags (Su and Yajima, 2015; version 0.5-7). Vague,

independent priors were derived from normal prior distributions (mean = 0 and standard

deviation = 1000) for the community-level habitat use and detection covariates. The posterior

distribution was obtained by running 3 chains of The Markov Chain Monte Carlo (MCMC),

with 140 000 iterations, a burn-in of 100 000 iterations and a thinning rate of 40. Therefore,

final inferences were made from a sample size of 3 000, and deemed adequate based off later

(

i statistics (Gelman and Rubin, 1992). In some settings, thinning does not help with

convergence and could be inefficient (Link and Eaton, 2012). Unfortunately, high

autocorrelation in my model was unavoidable; thus, requiring very long chains. With multiple

nodes being monitored, computer memory and storage was a major limitation, and

therefore, resulting in a high burn-in and thinning rate to reduce this cost (Broms, Hooten and

Fitzpatrick, 2016). Additionally, due to substantial post-processing required, whereby derived

parameters were calculated for each sampled value of the Markov chain, posing a substantial

32

computational burden, it was thought that overall results would be improved by reducing

autocorrelation in the chains being used through increased thinning (Link and Eaton, 2012).

2.6.2 | Candidate models

To avoid the risk of model over-parameterization that could reduce the precision of the

habitat use estimates, only a few hypotheses driven models were considered. These models

explored three broad drivers of mesocarnivore habitat use, namely:

1. Environmental drivers: NDVI and TRI

2. Interspecific drivers: Competition/predation from apex predators and domestic dogs

3. Anthropogenic drivers: Distance to PA edge and domestic dogs

A global model, incorporating all the above covariates, was also examined. Due to the survey

being designed for monitoring leopards, and leopards being detected across all surveyed PAs,

a second set of global and competition models were also considered, exploring the influence

of leopard relative abundance rather than just apex predator abundance as a whole. Finally,

a null model, which only incorporated detection covariates, was considered.

2.6.3 | Model fit assessment

Convergence was assessed visually by inspecting the produced MCMC chains and using the

Gelman-Rubin statistic, where

(

i

j 5k5

was considered to be acceptably converged (Gelman

and Rubin, 1992; Brooks and Gelman, 1998). “Lack of fit” of the data was evaluated using the

Bayesian p-value (Kéry, 2010), a posterior-predictive check that estimates the level of

dispersion in the data relative to the model. The Bayesian p-value is defined as the probability

that the replicated data are more extreme than the observed data. The Bayesian p-value was

estimated based on the Pearson’s

lO

discrepancy for binomial data, such that

DE%3lmno

OG

lo0p

O6

. The simulated data were assumed to be “perfect”, and thus allowing the Pearson

residuals to represent the fit of the model when all model assumptions were perfectly met

(Kéry, 2010). This created a fit metric that was equal to one when the Pearson residual was

greater for the observed data than the simulated data, or equal to zero otherwise. Therefore,

the final Bayesian p-value was calculated as the mean of the posterior sample of the model

fit metric, where a mean of 0.5 indicated perfect fit, and values between 0.05 and 0.95

33

indicated adequate model fit (Soto-Shoender et al., 2018). Finally, a “lack-of-fit” statistic was

reported

qrst

u

qtvw

u

, which was expected to equal 1 when the data fit the model perfectly (Kéry and

Schaub, 2012).

2.6.4 | Derived parameters

Many occupancy models can estimate the number of species in a community that were

unobserved during the sampling period, but this often requires prior knowledge of the total

number of species that can possibly occur at a site. However, there was no well-documented

mesocarnivore species list for the PAs considered in my study. Due to this, I could not derive

absolute species richness, but instead focused on comparing the relative species richness of

focal mesocarnivore communities in the different PAs.

The probability that a species was present during a survey, but not detected, was calculated

using MCMC algorithms, where

/01 4 5

for each iteration if the species was detected during

the survey. If the species was not detected, then a mean of

/01

was taken over all iterations.

This produced a probability that the species was present but overlooked, while also taking

into account the probability of habitat use of the species, PA, sampling effort and detection

probability across surveys.

The total number of species present during each survey

8

, or estimated species richness

3M16

,

was calculated as:

M14

V

/01

0

Station-level species richness was then estimated using a similar occurrence matrix as above,

adapted to the station-level, where species of interest

2

was present

3@0A 4 56

or absent

3@0A 4 76

in the area around station

?

. Species richness per station was calculated as

@0A/01

.

2.7 | Species richness - Generalized Additive Model

Generalized Additive Models (GAMs) are flexible models and can be used to maximize the

predictive quality of a covariate from various distributions, and thus allowing for better fit in

the presence of non-linear relationships and significant noise in the predictor covariates (Hill

34

and Lewicki, 2007). GAMs are often used to predict the likelihood of species presence and

abundance using environmental variables and have been shown to perform better than other

types of ecological predictive models (Guisan, Edwards and Hastie, 2002; Moisen and

Frescino, 2002). In my study, GAMs were applied to model the relationship between station-

level species richness and various predictor covariates.

General GAMs are expressed as:

a

3

x

6

4 d -

V

y0

3

`0

6

-z

]

0^e

where

a

3

x

6 is the link function defining the relationship between the response variable

(species richness) and the predictor variables (selected covariates).

d

is the intercept term,

<%

the number of covariates,

y0

is the spline smoothing function of each predictor

`0

, and

z

is

the residual error term (Wood, 2017). Thin plate regression splines were used as the

smoothing function for all covariates. This approach allowed lower ranked smooths to be

nested within higher ranked smooths, thus allowing conventional hypothesis testing methods

to be used to compare GAMs (Wood, 2003). Four GAMs were developed based on the

candidate models outlined for the RN model, namely, 1) Environmental: NDVI and TRI, 2)

Anthropogenic: Distance to PA edge and domestic dog abundance, and 3) Competition:

domestic dog and leopard abundance. The global GAM was based on the best-fit RN model

(Global B, Table 3.3), i.e., 4) Global: NDVI, TRI, distance to edge, dogs and leopard abundance.

The best-fitting GAM was selected based on the Akaike’s Information Criterion (AIC; Sagarese

et al., 2014). A fifth GAM was also developed using the second best-fit RN model (Global A;

Table 3.3). All GAMs were run using the R package ‘mgcv’ (Wood, 2011; version 1.8-28).

2.8 | Species richness - Variance model

In order to assess the effects of each variable on the variation in species richness, I used a log-

linear regression of the variance in species richness on all covariates (Kéry, 2010; Li, Bleisch

and Jiang, 2018). Species richness for station

2

was drawn from a normal distribution with

mean (

x0

) and variance

35f

{

6

.

35

The variance 3

l

6 of the

%`

covariates (COV), namely, NDVI, TRI, distance to PA edge, dog and

leopard abundance, was then calculated using:

QRS

3

f0

6

4 lU-

V

lWXYZW

]^_

W

Where

f0

is the variance in species richness for station

2

, and

lU

is the mean variance drawn

from a normal distribution with mean of zero and a variance of 0.001.

This modelling process was also carried out in R2Jags (Su and Yajima, 2015; version 0.5-7),

using uninformative priors derived from normal prior distributions with mean zero and

variance 0.001. The posterior distribution was obtained by running 3 chains of MCMC with 20

000 iterations, a burn-in of 10 000 iterations and a thinning rate of 10.

36

3 | RESULTS

3.1 | Descriptive results

A total of 392 642 photographs were collected over 13 823 trap nights for the seven camera

trap surveys (Table 3.1). The number of trap days were similar (1689

|

230 days) across all

surveys. Duplicates, ‘blank’ photos and photos of the research team accounted for 137 966

of the photographs and were removed prior to analyses. Independent captures were broken

down into a total of 64 species across 14 orders (including human unknown and “other”

photographs; Table 3.2). ZRR produced captures of the greatest number of species, whilst

Tembe the least (Table 3.1).

The total number of detections for the five mesocarnivores species varied markedly across

the different PAs (Table 3.2). Honey badger was the only species detected across all surveyed

PAs, while serval was detected in five PAs, side-striped jackals in four, black-backed jackals in

three and caracal in two of the PAs. When comparing PAs, ZRR was the only PA in which all

five mesocarnivores were detected (Table 3.2). The highest number of mesocarnivore

detections were recorded for ZRR (N = 117). Detections of black-backed jackal, honey badger

and serval were distributed evenly throughout ZRR (Figure 3.4). By contrast, Tembe had the

lowest number of detections across all focal species with honey badger being the only

mesocarnivore detected (N = 8). These detections were mostly confined to the southern

regions of Tembe (Figure 3.2). Eastern Shores had the highest number of a single

mesocarnivore species, honey badger, detected (N = 62; Table 3.2). These detections

occurred mostly along the eastern boundary of the PA, close to the ocean (Figure 3.1; Figure

2.1). The highest number of detections for side-striped jackal occurred in uMkhuze and

appeared to be evenly spread throughout the surveyed area (Figure 3.3). HiP had the second

lowest total number of mesocarnivore detections (N = 10), however these were spread over

three species, namely honey badger, side-striped jackal and serval (Table 3.2). All detections

of side-striped jackal and serval in HiP occurred at a single camera station (Figure 3.3).

37

Table 3.1 Summary of the camera trap surveys conducted in seven PAs in KZN, South Africa,

namely Eastern Shores Section of iSimangaliso Wetland Park (E.Shores), Hluhluwe-Imfolozi

Park (HiP), Ithala Game Reserve, Somkhanda Game Reserve, Tembe Elephant Park, uMkhuze

Game Reserve and Zululand Rhino Reserve (ZRR). Area is the total area (km2) covered by the

camera trap stations. Trap days is the total number of days cameras were active at each site,

independent captures refers to the total number of independent photographs (

!

30 min) of

target species and total species is a count of the total number of different animal species

(domestic and wildlife) detected by that camera trap survey.

PA

Area

No. of

stations

Total trap

days

Total no. of

captures

Independent

captures

Total

species

E.Shores

148

41

1834

114705

83169

41

HiP

336

46

1957

93495

62438

44

Ithala

236

31

1349

36864

21065

45

Somkhanda

229

40

1739

23310

13460

44

Tembe

166

32

1392

42714

21863

40

uMkhuze

146

40

1759

51524

36856

45

ZRR

200

40

1793

30030

15825

48

38

Table 3.2. Summary of the number of captures of the 64 species detected for 1-day occasion periods (i.e., before pooling) across the seven PAs

in KZN, South Africa, namely Eastern Shores Section of iSimangaliso Wetland Park (E.Shores), Hluhluwe-Imfolozi Park (HiP), Ithala Game Reserve,

Somkhanda Game Reserve, Tembe Elephant Park, uMkhuze Game Reserve and Zululand Rhino Reserve (ZRR).

Species

Common name

E.Shores

HiP

Ithala

Somkhanda

Tembe

uMkhuze

ZRR

Carnivora

Acinonyx jubatus

Cheetah

0

1

0

19

14

Aonyx capensis

Cape Clawless Otter

0

1

0

Atilax paludinosus

Water Mongoose

45

1

4

3

6

1

Canis adustus

Side-striped Jackal

13

1

0

35

11

Canis mesomelas

Black-backed Jackal

0

5

10

0

46

Caracal caracal

Caracal

0

8

0

4

Crocuta crocuta

Spotted Hyaena

334

407

6

38

1

165

34

Felis serval

Serval

21

3

26

11

0

34

Galerella sanguinea

Slender Mongoose

0

2

14

10

3

1

Genetta tigrina

Large-spotted Genet

62

31

57

56

110

100

38

Hyaena brunnea

Brown Hyaena

0

75

5

0

43

Ichneumia albicauda

White-Tailed Mongoose

4

43

42

54

44

103

88

Ictonyx striatus

Striped Polecat

0

15

0

Lycaon pictus

Wild Dog

0

109

0

154

45

147

50

Mellivora capensis

Honey Badger

62

6

14

5

8

11

22

Mungos mungo

Banded Mongoose

9

0

2

7

2

Panthera leo

Lion

0

124

0

197

49

118

Panthera pardus

Leopard

217

121

359

58

102

115

88

Rhynchogale melleri

Meller's Mongoose

0

1

0

1

0

Lagomorpha

Lepus saxatalis

Scrub Hare

1

69

96

18

40

260

161

39

Table 3.2. (Continued)

Species

Common name

E.Shores

HiP

Ithala

Somkhanda

Tembe

uMkhuze

ZRR

Perissodactyla

Ceratotherium simum

White Rhinoceros

24

582

123

74

28

258

189

Diceros bicornis

Black Rhinoceros

11

51

66

20

6

37

36

Equus quagga

Plains Zebra

245

446

947

298

32

510

434

Primates

Cercopithecus albogularis

Samango Monkey

71

3

0

48

0

Cercopithecus pygerythus

Vervet Monkey

350

54

468

1295

44

386

248

Otolemur crassicaudatus

Greater Bushbaby

0

3

0

2

1

Papio ursinus

Chacma Baboon

88

186

1281

3

0

333

35

Proboscidae

Loxodonta africana

African Elephant

8

609

207

70

634

75

164

Rodentia

Hystrix africaeaustralis

Cape Porcupine

232

30

149

205

123

104

278

Thryonomys swinderianus

Cane Rat

0

11

2

8

0

4

2

Ruminantia

Aepyceros melampus

Impala

37

169

422

1570

595

2809

1562

Alcelaphus buselaphus

Red Hartebeest

0

2

1

0

Cephalophus natalensis

Red Duiker

1441

15

1

132

398

167

131

Connochaetes taurinus

Blue Wildebeest

43

72

267

601

61

708

540

Damaliscus lunatus

Tsessebe

0

1

0

Giraffa camelopardalis

Giraffe

0

349

262

150

215

238

552

Kobus ellipsiprymnus

Waterbuck

656

39

101

39

6

0

88

Neotragus moschatus

Suni

0

13

5

0

Oreatragus oreotragus

Klipspringer

0

3

0

40

Table 3.2. (Continued)

Species

Common name

E.Shores

HiP

Ithala

Somkhanda

Tembe

uMkhuze

ZRR

Ruminantia

Potamochoerus porcus

Bushpig

77

19

68

43

1

31

13

Raphicerus campestris

Steenbok

0

4

0

10

0

Redunca arundinum

Common Reedbuck

30

0

1

0

1

0

32

Redunca fulvorufola

Mountain Reedbuck

0

1

19

0

22

Sylvicapra grimmia

Common Duiker

26

99

30

130

207

238

95

Syncerus caffer

African Buffalo

340

407

9

189

4

57

196

Taurotragus oryx

Eland

0

48

0

Tragelaphus angasii

Nyala

72

332

98

1110

3037

2290

2268

Tragelaphus scriptus

Bushbuck

719

14

323

182

18

2

46

Tragelaphus strepsiceros

Kudu

736

82

765

393

72

180

582

Squamata

Varanus species

Monitor lizard

1

0

Suiformes

Hippopotamus amphibious

Hippopotamus

639

8

0

39

7

145

14

Phacochoerus africanus

Warthog

539

382

353

2194

70

862

1901

Testudines

Geochelone pardalis

Leopard Tortoise

0

1

0

Tubulidentata

Orycteropus afer

Aardvark

70

7

29

93

0

19

44

41

Table 3.2. (Continued)

a Excluding research team

Species

Common name

E.Shores

HiP

Ithala

Somkhanda

Tembe

uMkhuze

ZRR

Domestic

Bos taurus

Cow

0

299

0

Canis familiaris

Dog

8

1

6

36

1

15

32

Capra aegagrus

Goat

0

1

Equus ferus

Horse

3

1

0

Felis catus

Cat

1

0

1

0

Human

Homo sapien

Humana

2077

254

568

545

153

268

242

Vehicle

73620

57032

13384

2884

15366

25843

5444

Other

Bat species

1

3

0

1

8

6

Bird species

155

66

197

293

80

115

257

Insect species

19

25

38

31

2

0

3

Unknown

7

169

45

31

0

98

0

42

Figure 3.1 Mesocarnivore capture frequencies recorded at camera trap stations in 1) Eastern Shores Section of iSimangaliso Wetland Park (Eastern Shores) and

3) Ithala Game Reserve during 2015 surveys. Numbers correspond to survey number in Figure 2.1. Points represent station locations scaled by the number of

independent captures (counts) of that species.

0 5 10km

Honey badger

Serval

Side−striped jackal

Count

0

2

4

6

8

0 3 6km 0 3 6km 0 3 6km

Honey badger Serval Black−backed jackal

Count

0

1

2

3

4

0 5 10km 0 5 10km 0 5 10km

Honey badger Serval Side−striped jackal

Count

0

2

4

6

8

0 3 6km

Honey badger

Serval

Black−backed jackal

Count

0

1

2

3

4

111

333

0 5 10km

Honey badger Serval Side−striped jackal

Count

0

2

4

6

8

0 3 6km 0 3 6km 0 3 6km

Honey badger Serval Black−backed jackal

Count

0

1

2

3

4

0 5 10km

Honey badger Serval Side−striped jackal

Count

0

2

4

6

8

0 3 6km 0 3 6km 0 3 6km

Honey badger Serval Black−backed jackal

Count

0

1

2

3

4

0 5 10km

Honey badger Serval Side−striped jackal

Count

0

2

4

6

8

0 3 6km 0 3 6km 0 3 6km

Honey badger Serval Black−backed jackal

Count

0

1

2

3

4

43

Figure 3.2 Mesocarnivore capture frequencies recorded at camera trap stations in 4) Somkhanda Game Reserve and 5) Tembe Elephant Park during 2015

surveys. Numbers correspond to survey number in Figure 2.1. Points represent station locations scaled by the number of independent captures (counts) of that

species.

0 4 8km

Serval

Caracal

Honey badger

Black−backed jackal

Count

0

1

2

3

4

5

4 4 4

4

0 4 8km 0 4 8km

Black−backed jackal Honey badger

Count

0

1

2

3

4

5

0 3 6km

Honey badger

Count

0.0

0.5

1.0

1.5

2.0

0 4 8km

0 4 8km 0 4 8km

Serval

Caracal Honey badger Black−backed jackal

Count

0

1

2

3

4

5

44

Figure 3.3 Mesocarnivore capture frequencies recorded at camera trap stations in 2) Hluhluwe-Imfolozi Park (HiP) and 6) uMkhuze Game Reserve during 2015

surveys. Numbers correspond to survey number in Figure 2.1. Points represent station locations scaled by the number of independent captures (counts) of that

species.

0 5 10km

Honey badger

Side−striped jackal

Count

0

1

2

3

4

5

0 5 10km

Honey badger

Side−striped jackal

Serval

Count

0

1

2

3

222

6 6

45

Figure 3.4 Mesocarnivore capture frequencies recorded at camera trap stations in 7) Zululand Rhino Reserve (ZRR) during 2015 surveys. Numbers correspond

to survey number in Figure 2.1. Points represent station locations scaled by the number of independent captures (counts) of that species.

0 3 6km

Caracal

Serval

Black−backed jackal

Side−striped jackal

Honey badger

Count

0

2

4

6

0 3 6km

Caracal Serval

Black−backed jackal Side−striped jackal Honey badger

Count

0

2

4

6

0 3 6km

Caracal Serval

Black−backed jackal Side−striped jackal Honey badger

Count

0

2

4

6

0 3 6km

Caracal Serval

Black−backed jackal Side−striped jackal Honey badger

Count

0

2

4

6

0 3 6km

Caracal Serval

Black−backed jackal Side−striped jackal Honey badger

Count

0

2

4

6

0 3 6km

Caracal Serval

Black−backed jackal Side−striped jackal Honey badger

Count

0

2

4

6

7 7 7

77

46

3.2 | Model selection

Seven hypothesis driven candidate models were considered (Table 3.3). These models

included two competition models which incorporated species interaction covariates, an

anthropogenic model incorporating human disturbance, an environmental model

incorporating habitat variables, two global models and a null model. Detection covariates, PA

and trail, remained the same for all candidate models; thus, differences in model fit were due

to occupancy covariates. Goodness of fit tests showed that model Global B, which included

all of the predictor variables and leopard abundance, provided the best fit for the observed

data, with the Bayesian p-value closest to 0.5 (p = 0.628) and a lack of fit statistic closest to 1

(lack of fit = 1.060; Table 3.3). All RN models had a mean

!

" below 1.1, showing that even the

model with the least fit, i.e., the lowest ranked model “Competition A”, still had acceptable

convergence.

Table 3.3 Summary of Royle-Nichol (RN) multi-species models ordered by decreasing model

fit based on each model’s Bayesian p-value, lack of fit statistic and mean Gelman-Rubin

statistic (

!

"). Model covariates included NDVI (Normalized Difference Vegetation Index), TRI

(Terrain Ruggedness Index), Dist2edge (distance to the edge of the PA), Dogs (domestic dog

relative abundance), Leopards (leopard relative abundance), PA (survey site), and trail

(feature along which the camera was placed).

Models

Bayesian

p-value

Lack of fit

Mean #

$

Global B

%(NDVI+TRI+Dist2edge+Dogs+Leopards+PA)

&(PA+Trail)

0.628

1.060

1.020

Global A

%(NDVI+TRI+Dist2edge+Dogs+Apex+PA)

&(PA+Trail)

0.637

1.067

1.017

Anthropogenic

%'Dist2edge+Dogs+PA)(&(PA+Trail)

0.644

1.070

1.021

Competition B

%(Dogs+Leopards+PA)(&(PA+Trail)

0.681

1.088

1.009

Null

%(PA) &(PA+Trail)

0.684

1.086

1.013

Environmental

%(NDVI+TRI+PA)(&(PA+Trail)

0.686

1.077

1.011

Competition A

%(Dogs+Apex+PA)(&(PA+Trail)

0.717

1.098

1.012

3.3 | Mesocarnivore detections and habitat use

There was limited variation in detection probability of the five species across the PAs (Figure

3.5), with most probabilities falling below 0.1. The highest detection probabilities were

obtained for black-backed jackals in Somkhanda and serval in HiP (Figure 3.5) with detection

47

for both species localized to small regions of each PA (Figure 3.2 and 3.3 respectively). This

translated to lowered probabilities of habitat use for serval and black-backed jackal in both

PAs (Figure 3.5). Habitat use varied more between species and PAs than detection probability,

with 30% of the mesocarnivore species having a probability of use greater than 0.5 (Figure

3.5). The majority of mesocarnivore species in Tembe had low estimated habitat use (< 0.03).

ZRR, on the other hand, had relatively high habitat use estimates for all five detected species,

with caracal and side-striped jackal greater than 0.3 and the others all above 0.65. Honey

badger and serval had the highest number of detections across the PAs, which translated to

a higher probability of habitat use (Figure 3.6). Black-backed jackals had the highest

probability of detection. Though it is important to note that there were no significant

differences in detection or probability of habitat use between species (Figure 3.6). The

likelihood of a mesocarnivore species being present, but overlooked by the camera survey,

was generally low (

)(

» 0.2, Table 3.4), with the highest probability for side-striped jackals in

Ithala (

)(

= 0.4).

48

Figure 3.5 Species-specific detection and habitat use estimates per PA. Distribution of the

total number of detections for 5-day pooled data (N), mean per-individual detection

probabilities (R) and mean use (

%

). PAs included Eastern Shores Section of iSimangaliso

Wetland Park (E.shores), Hluhluwe-Imfolozi Park (HiP), Ithala Game Reserve, Somkhanda

Game Reserve, Tembe Elephant Park, uMkhuze Game Reserve and Zululand Rhino Reserve

(ZRR). Error bars show 95% Bayesian credible intervals.

No. of detections (N)

Detection probability (R)

Habitat use (Y)

E.shores

HiP

Ithala

Somkhanda

Tembe

uMkhuze

ZRR

0

20

40

60

0.0

0.1

0.2

0.3

0.4

0.5

0.00

0.25

0.50

0.75

1.00

Side−striped jackal

Serval

Honey badger

Caracal

Black−backed jackal

Side−striped jackal

Serval

Honey badger

Caracal

Black−backed jackal

Side−striped jackal

Serval

Honey badger

Caracal

Black−backed jackal

Side−striped jackal

Serval

Honey badger

Caracal

Black−backed jackal

Side−striped jackal

Serval

Honey badger

Caracal

Black−backed jackal

Side−striped jackal

Serval

Honey badger

Caracal

Black−backed jackal

Side−striped jackal

Serval

Honey badger

Caracal

Black−backed jackal

49

Figure 3.6 Species-specific detection and habitat use estimates across all PAs. Distribution of

the total number of detections for 5-day pooled data (N), mean per-species detection

probabilities (R) and mean use (

%

) across all seven PAs in KZN, namely Eastern Shores Section

of iSimangaliso Wetland Park (E.shores), Hluhluwe-Imfolozi Park (HiP), Ithala Game Reserve,

Somkhanda Game Reserve, Tembe Elephant Park, uMkhuze Game Reserve and Zululand

Rhino Reserve (ZRR). Error bars show 95% Bayesian credible intervals.

Table 3.4 Probability of presence

')*

for each mesocarnivore species for each PA in KZN.

)

= 1.000 indicates that the species was detected by the camera survey,

)

< 1.00 is the

probability of the species being present but overlooked by the camera survey.

None of the modelled covariates had a significant effect on community-level (Figure 3.7) or

species-level habitat use (Figure 3.8; BCI overlap zero). TRI, leopard abundance and domestic

dog abundance all had a negative effect on mesocarnivore habitat use and had the narrowest

BCIs (Figure 3.7), whilst NDVI and distance to edge had a positive influence on mesocarnivore

use of PAs.

Species-level response to the different covariates varied markedly (Figure 3.8). Serval and

caracal habitat use were positively associated with distance to PA edge. Serval and side-

striped jackal habitat use increased with higher domestic dog abundance. Honey badger had

Species

Eastern

Shores

HiP

Ithala

Somkhanda

Tembe

uMkhuze

ZRR

Black-backed jackal

0.079

0.117

1.000

0.172

0.137

1.000

Caracal

0.084

0.135

0.212

1.000

0.173

0.193

1.000

Honey badger

1.000

Serval

1.000

0.202

0.186

1.000

Side-striped jackal

1.000

0.395

0.287

0.245

1.000

No. of detections (N)

Detection probability (R)

Habitat use (Y)

0 50 100 0.0 0.1 0.2 0.3 0.00 0.25 0.50 0.75 1.00

Caracal

Side−striped jackal

Black−backed jackal

Serval

Honey badger

50

the largest positive association with leopard abundance, whilst black-backed jackals had the

largest negative association with leopard abundance.

Figure 3.7 Influence of standardized hyperparameters on community-level mesocarnivore

habitat use of PAs in KZN, based on the RN model. Hyperparameters included NDVI

(Normalized Difference Vegetation Index), TRI (Terrain Ruggedness Index), Distance to the

edge of the PA, Dog abundance (Domestic dog relative abundance), and Leopard relative

abundance. Points indicate the posterior mean, and lines give the 95% Bayesian credible

intervals. PA was also included as a habitat use hyperparameter but BCI range was

exceedingly large (-18.885 to 17.890), and thus was not included in this graph. Additionally,

detection hyperparameters Trail (feature along which the camera was placed) and PA (survey

site) were tested but BCI ranges were also exceedingly large (-18.701 to 17.925 and -18.885

to 17.890 respectively), and thus were not included in this graph.

Leopard abundance

Dog abundance

Distance to edge

TRI

NDVI

−1.0

−0.5

0.0

0.5

1.0

Beta coefficient

51

Figure 3.8 Influence of covariates on species-specific habitat use. Standardized beta

coefficients and 95% Bayesian credible intervals for the influence of A) Protected Area (PA),

B) Normalized Difference Vegetation Index (NDVI), C) Terrain Ruggedness Index (TRI), D)

Distance to the edge of the PA, E) Domestic dog relative abundance and F) Leopard relative

abundance on mesocarnivore species use of the area around the camera stations across PAs

in KZN, based on the RN model.

3.4 | PA mesocarnivore species richness

Overall, mesocarnivore species richness varied between the different PAs (Figure 3.9). The

highest species richness was estimated for ZRR, where all five (100%) mesocarnivore species

were detected (Figure 3.9), while Tembe had the lowest species richness with a mean species

richness of 1.79 (36%). This pattern was also seen in the total number of species recorded

(Table 3.1), with ZRR having the highest number of species (48), and Tembe having the lowest

(40). Across all the PAs, the mean estimated species richness was higher than the total

−4

−2

0

2

PA

A

−1.5

−1.0

−0.5

0.0

0.5

1.0

NDVI

B

−1.0

−0.5

0.0

0.5

1.0

TRI

C

−1.0

−0.5

0.0

0.5

1.0

Distance to

edge

D

−0.2

−0.1

0.0

0.1

0.2

Dog abundance

E

−0.10

−0.05

0.00

0.05

0.10

Leopard abundance

F

−0.10

−0.05

0.00

0.05

0.10

Black−backed jackal

Caracal

Honey badger

Serval

Side−striped jackal

Leopards

−0.10

−0.05

0.00

0.05

0.10

Black−backed jackal

Caracal

Honey badger

Serval

Side−striped jackal

Leopards

52

number of detected species (Figure 3.9), revealing that the model took into account individual

species detection probabilities. Somkhanda had a significantly higher estimated species

richness compared to that of Eastern Shores, HiP, Tembe and uMkhuze (no BCI overlap).

Tembe had a significantly lower estimated species richness relative to HiP, Ithala, Somkhanda,

and ZRR. Species richness estimates were less precise, that is exhibited greater standard

deviations and wider BCIs, for surveys with lower survey effort, i.e., Tembe and Ithala (Table

2.1).

Figure 3.9 Estimated PA-level mesocarnivore species richness. Points represent estimated

mean species richness and error bars represent 95% Bayesian credible intervals around

estimated means. Red numbers represent the total number of detected mesocarnivore

species within that PA.

3.5 | Covariates influencing mesocarnivore species richness

The best fitting GAM (Table 3.5) was Global B, which had the lowest AIC (642.53) and the

highest explained deviance, accounting for 30.5% of the deviance shown in the data. NDVI,

distance to the edge of the PA and leopard abundance all had statistically significant effects

on mesocarnivore species richness (Figure 3.10; Table 3.5). Partial response plots showed

species richness being greater at lower NDVI values, followed by a U-shaped dip in species

richness for increasing NDVI (Figure 3.10). Distance to the edge of the PA had the clearest

3 3

3 4

1

2

5

1

2

3

4

5

Eshores

HiP

Ithala

Somkhanda

Tembe

uMkhuze

ZRR

Mean species richness

53

relationship with species richness, with the estimated number of mesocarnivore species

increasing closer to the boundaries of the PA. Although leopard abundance as a predictor

variable provided a statistically good fit to the data, it was difficult to interpret the complex

relationship with mesocarnivore species richness. TRI did not have a significant effect on

mesocarnivore richness, and domestic dogs showed an extremely variable effect, with the

confidence bands being too wide to make any inferences (Figure 3.10).

A different GAM was also produced for the second best-fit model (Global A, Table 3.3), but

only accounted for 25.6% of the deviance shown in the data (Table 3.6). NDVI, TRI, distance

to PA edge and apex predator relative abundance all had statistically significant effects on

mesocarnivore species richness (Table 3.6; Figure 3.11). Relationships between species

richness and NDVI, TRI and distance to PA edge showed similar associations seen in the

original GAM, based on Global B (Figure 3.10). Apex predator abundance appeared to have a

quadratic relationship with mesocarnivore species richness, with an initial decline followed

by a gradual increase in richness with greater abundance of apex predators (Figure 3.11).

Finally, variation in species richness was significantly influenced by NDVI and domestic dog

abundance (Table 3.7). Both variables showed inverse relationships with species richness

variance, with increased NDVI and dog abundance resulting in reduced variability in station-

level mesocarnivore richness estimates, or a more homogenous richness landscape.

54

Figure 3.10 Generalised Additive Model (GAM) plots, based on Global B (Table 3.3 and Table

3.5), showing the partial effects of selected covariates on mesocarnivore species richness in

PAs in KZN. The x-axis is the range of the specified covariate, with the tick marks representing

locations of observed data points. The y-axis is the additive contribution of the covariate to

the non-parametric GAM smoothing function. Grey shaded areas indicate the 95% confidence

intervals.

3000 4000 5000 6000 7000 8000

0.2 0.4 0.6 0.8

NDVI

s(NDVI)

5 10 15 20 25

0.2 0.4 0.6 0.8

TRI

s(TRI)

0 2000 6000 10000

0.2 0.4 0.6 0.8

Distance to edge (m)

s(Distance to edge)

0 5 10 15 20 25 30

0.2 0.4 0.6 0.8

Dog abundance

s(Dog abundance)

0 10 20 30 40 50

0.2 0.4 0.6 0.8

Leopard abundance

s(Leopard abundance)

55

Table 3.5 Estimated GAM coefficients ordered by best-fitting model based on Akaike’s

Information Criterion (AIC) and percentage deviance explained (Dev%). Global model based

on Global B (Table 3.3). Estimated degrees of freedom (Edf), F-statistic (F), and probability

level of significance (P) also provided. P-value codes: “***”P<0.001, “**”P<0.01, “*”P<0.05,

and “ ”P>0.05. Model covariates included NDVI (Normalized Difference Vegetation Index), TRI

(Terrain Ruggedness Index), Dist2edge (distance to the edge of the PA), Dogs (Domestic dog

relative abundance) and Leopards (leopard relative abundance).

Model and covariates

Edf

F

P

AIC

Dev(%)

Global

s(NDVI)

4.197

5.848

***

642.528

30.5

s(TRI)

3.129

0.572

s(Dist2edge)

1.624

2.312

***

s(Dogs)

5.701

1.255

s(Leopards)

6.474

1.962

**

Environmental

s(NDVI)

4.198

4.344

***

668.466

15.1

s(TRI)

2.983

0.699

Anthropogenic

s(Dist2edge)

5.242

3.058

***

682.608

9.27

s(Dogs)

4.464e-10

0.000

Competition

s(Dogs)

5.016e-09

0.000

696.903

5.26

s(Leopards)

6.547

1.345

56

Figure 3.11 Generalised Additive Model (GAM) plots, based on Global A (Table 3.3 and Table

3.5), showing the partial effects of selected covariates on mesocarnivore species richness in

PAs in KZN. The x-axis is the range of the specified covariate, with the tick marks representing

locations of observed data points. The y-axis is the additive contribution of the covariate to

the non-parametric GAM smoothing function. Grey shaded areas indicate the 95% confidence

intervals.

3000 4000 5000 6000 7000 8000

0.2 0.3 0.4 0.5 0.6 0.7

NDVI

s(NDVI)

5 10 15 20 25

0.2 0.3 0.4 0.5 0.6 0.7

TRI

s(TRI)

0 2000 6000 10000

0.2 0.3 0.4 0.5 0.6 0.7

Distance to edge (m)

s(Distance to edge)

0 5 10 15 20 25 30

0.2 0.3 0.4 0.5 0.6 0.7

Dog abundance

s(Dog abundance)

0 10 20 30 40

0.2 0.3 0.4 0.5 0.6 0.7

Apex abundance

s(Apex abundance)

57

Table 3.6 Estimated GAM coefficients based on Global A (Table 3.3). Estimated degrees of

freedom (Edf), F-statistic (F), probability level of significance (P), and percentage deviance

explained (Dev%) all provided. P-value codes: “***”P<0.001, “**”P<0.01, “*”P<0.05, and “

”P>0.05. Model covariates included NDVI (Normalized Difference Vegetation Index), TRI

(Terrain Ruggedness Index), Dist2edge (distance to the edge of the PA), Dogs (Domestic dog

relative abundance) and Apex (apex predator relative abundance).

Table 3.7 Mean estimates, standard deviation and 95% Bayesian credible intervals of the

covariates hypothesized to influence the variance in station-specific (point-level) species

richness. Values in bold indicate covariates for which the credible intervals (95% BCI) did not

overlap zero. Model covariates included NDVI (Normalized Difference Vegetation Index), TRI

(Terrain Ruggedness Index), Dist2edge (distance to the edge of the PA), Dogs (domestic dog

relative abundance), and Leopards (leopard relative abundance). Mean Gelman-Rubin

statistic (

!

") also reported.

Parameters

Mean

Standard deviation

95% BCI

#

$

+,

Mean richness

-0.238

0.135

-0.498

0.028

1.00

-,

Mean variance

1.317

0.058

1.278

1.429

1.00

./

NDVI

-0.335

0.091

-0.399

-0.150

1.00

-0

TRI

-0.166

0.108

-0.372

0.048

1.00

-1

Dist2edge

-0.011

0.103

-0.206

0.196

1.00

.2

Dogs

-0.082

0.034

-0.138

-0.008

1.00

-3

Leopards

-0.020

0.013

-0.046

-0.007

1.00

Model and covariates

Edf

F

P

Dev(%)

Global

s(NDVI)

4.258

3.896

***

25.6

s(TRI)

3.329

0.902

*

s(Dist2edge)

4.184

2.491

***

s(Dogs)

6.177e-07

0.000

s(Apex)

1.796

0.037

*

58

4 | DISCUSSION

Mesocarnivores persist across diverse human-modified ecosystems including agricultural

(Drouilly, Clark and O’Riain, 2018), peri-urban (Serieys et al., 2019), industrial (Loock et al.,

2018) and protected (Tambling et al., 2018) landscapes. Given the extent of human-wildlife

conflict in agricultural and peri-urban areas, PAs are often presumed to provide an essential

refuge for wildlife including mesocarnivores. Yet this may not be the case, as PAs face a wide

variety of threats and include a range of management practices that can reduce their

conservation potential (Balme, Slotow and Hunter, 2010; Watson et al., 2014; Santini et al.,

2016) and even actively persecute mesocarnivores such as black-backed jackal (Nattrass and

Conradie, 2015; Minnie, Gaylard and Kerley, 2016). My study investigated mesocarnivore

habitat use and species richness across seven PAs in KZN, South Africa, and revealed that not

only was

Mesocarnivores in Protected Areas: ecological and anthropogenic determinants of habitat use in northern Kwa-Zulu Natal, South Africa

Abstract